Project Overview¶
From a high level vantage, this is an image classification project.
With image classification, the input is an image and the output is a class.
Here, the problem domain is narrowed to the domain of dogs.
Specifically the intent is to input an image of a dog and output the dog breed.
This is a popular project of Udacity's, featured in many nanodegree programs.
The data set is curated by Udacity upon their project workspace.
This data set consists of over 8000 dog images from 133 breeds.
Additional input data includes uploaded image file aquired from a camera phone.
Problem Statement¶
Given a picture, identify if it contains a dog, and if so - reveal it's likely breed;
in the case no dog was found, search to see if it contains a human face and if so,
reveal a dog breed of similarity to the face; otherwise reveal that neither a dog
nor a face was found. Here an algorithm is developed to address this aim.A typical strategy to building classifiers is to use supervised learning.
Here, a data set of images with known class names is used to build a classification model.
The known data is split into train, validation, and test portions which are used to
optimize the performance of the classifier by tuning weight parameters in the model.
Various models are attempted, from simple to complex until a satisfactory result is found.
In this project a simple convolutional neural network is attempted, followed by the approach
of adopting pre-trained models tailored to classify dog breeds well. As the complexity
of the models increases, the performance improves, culminating in amazing results.
Model Metrics¶
In this work the main metric used is accuracy.
That is, when test data is run through a model,
accuracy is the percentage of correct predictions
out of all predictions.
Accuracy figures are typically given in percentage.
Attempts are made to optimize accuracy, within reason.
For Classifiers, Precision and Recall are also used.
These give a sense of false negatives and positives.
However in the context of Dogs this isn't critical.
An accurate Dog breed assessment is done by DNA testing.
Data Scientist Nanodegree¶
Convolutional Neural Networks¶
Project: Write an Algorithm for a Dog Identification App¶
This notebook walks you through one of the most popular Udacity projects across machine learning and artificial intellegence nanodegree programs. The goal is to classify images of dogs according to their breed.
If you are looking for a more guided capstone project related to deep learning and convolutional neural networks, this might be just it. Notice that even if you follow the notebook to creating your classifier, you must still create a blog post or deploy an application to fulfill the requirements of the capstone project.
Also notice, you may be able to use only parts of this notebook (for example certain coding portions or the data) without completing all parts and still meet all requirements of the capstone project.
In this notebook, some template code has already been provided for you, and you will need to implement additional functionality to successfully complete this project. You will not need to modify the included code beyond what is requested. Sections that begin with '(IMPLEMENTATION)' in the header indicate that the following block of code will require additional functionality which you must provide. Instructions will be provided for each section, and the specifics of the implementation are marked in the code block with a 'TODO' statement. Please be sure to read the instructions carefully!
In addition to implementing code, there will be questions that you must answer which relate to the project and your implementation. Each section where you will answer a question is preceded by a 'Question X' header. Carefully read each question and provide thorough answers in the following text boxes that begin with 'Answer:'. Your project submission will be evaluated based on your answers to each of the questions and the implementation you provide.
Note: Code and Markdown cells can be executed using the Shift + Enter keyboard shortcut. Markdown cells can be edited by double-clicking the cell to enter edit mode.
The rubric contains optional "Stand Out Suggestions" for enhancing the project beyond the minimum requirements. If you decide to pursue the "Stand Out Suggestions", you should include the code in this IPython notebook.
Why We're Here¶
In this notebook, you will make the first steps towards developing an algorithm that could be used as part of a mobile or web app. At the end of this project, your code will accept any user-supplied image as input. If a dog is detected in the image, it will provide an estimate of the dog's breed. If a human is detected, it will provide an estimate of the dog breed that is most resembling. The image below displays potential sample output of your finished project (... but we expect that each student's algorithm will behave differently!).
In this real-world setting, you will need to piece together a series of models to perform different tasks; for instance, the algorithm that detects humans in an image will be different from the CNN that infers dog breed. There are many points of possible failure, and no perfect algorithm exists. Your imperfect solution will nonetheless create a fun user experience!
The Road Ahead¶
We break the notebook into separate steps. Feel free to use the links below to navigate the notebook.
- Step 0: Import Datasets
- Step 1: Detect Humans
- Step 2: Detect Dogs
- Step 3: Create a CNN to Classify Dog Breeds (from Scratch)
- Step 4: Use a CNN to Classify Dog Breeds (using Transfer Learning)
- Step 5: Create a CNN to Classify Dog Breeds (using Transfer Learning)
- Step 6: Write your Algorithm
- Step 7: Test Your Algorithm
Step 0: Import Datasets¶
Import Dog Dataset¶
In the code cell below, we import a dataset of dog images. We populate a few variables through the use of the load_files function from the scikit-learn library:
train_files,valid_files,test_files- numpy arrays containing file paths to imagestrain_targets,valid_targets,test_targets- numpy arrays containing onehot-encoded classification labelsdog_names- list of string-valued dog breed names for translating labels
from sklearn.datasets import load_files
from keras.utils import np_utils
import numpy as np
import pandas as pd
from glob import glob
import os
import PIL
from PIL import Image
from sklearn.metrics import classification_report
import matplotlib.pyplot as plt
from matplotlib import image as mplim
%matplotlib inline
%config InlineBackend.figure_format='retina'
import seaborn as sb
sb.set_style("darkgrid")
# define function to load train, test, and validation datasets
def load_dataset(path):
'''
import a data path based upon udacity curated dataset
load data files using sklearn utility
return files and corresponding targets
'''
data = load_files(path)
dog_files = np.array(data['filenames'])
dog_targets = np_utils.to_categorical(np.array(data['target']), 133)
return dog_files, dog_targets
# load train, test, and validation datasets
train_files, train_targets = load_dataset('../../../data/dog_images/train')
valid_files, valid_targets = load_dataset('../../../data/dog_images/valid')
test_files, test_targets = load_dataset('../../../data/dog_images/test')
# load list of dog names
dog_names = [item[20:-1] for item in sorted(glob("../../../data/dog_images/train/*/"))]
# print statistics about the dataset
print('There are %d total dog categories.' % len(dog_names))
print('There are %s total dog images.\n' % len(np.hstack([train_files, valid_files, test_files])))
print('There are %d training dog images.' % len(train_files))
print('There are %d validation dog images.' % len(valid_files))
print('There are %d test dog images.'% len(test_files))
Import Human Dataset¶
In the code cell below, we import a dataset of human images, where the file paths are stored in the numpy array human_files.
import random
random.seed(8675309)
# load filenames in shuffled human dataset
human_files = np.array(glob("../../../data/lfw/*/*"))
random.shuffle(human_files)
# print statistics about the dataset
print('There are %d total human images.' % len(human_files))
Data Exloration and Visualization, Image Count by Breed¶
# Data Representation of Breed Counts
nnames = [s.split('.')[1] for s in dog_names]
train_breeds = np.asarray([nnames[np.argmax(train_targets[i])] for i in range(len(train_targets))])
valid_breeds = np.asarray([nnames[np.argmax(valid_targets[i])] for i in range(len(valid_targets))])
test_breeds = np.asarray([nnames[np.argmax(test_targets[i])] for i in range(len(test_targets))])
cnt_trn = []
cnt_val = []
cnt_tst = []
for name in nnames:
cnt_trn.append(np.sum(train_breeds==name))
cnt_val.append(np.sum(valid_breeds==name))
cnt_tst.append(np.sum(test_breeds==name))
cnt_sum = [sum(x) for x in zip(cnt_trn, cnt_val, cnt_tst)]
df_breed = pd.DataFrame.from_dict({'name': nnames,
'train': cnt_trn,
'valid': cnt_val,
'test': cnt_tst,
'total': cnt_sum,
})
df_breed = df_breed.sort_values(by='total', ascending=False)
df_breed = df_breed.drop(['total'], axis=1)
df_breed = df_breed.reset_index(drop=True)
df_breed.head()
# Median Image Count Dog Breed
df_breed.iloc[66]['name']
# Breed Count Statistics
df_breed.describe()
Image Count Reflection:¶
The test and validation sets are roughly similar.
The train set is roughly 8x the size of the others.
The breed with the most images contains ~3x more
than the breed with the least.
fig, ax1 = plt.subplots(1,1, figsize=(7, 12), dpi=120)
df_breed.plot.barh(x = 'name', stacked=True,
fontsize = 5, ax = ax1,
alpha = 0.5)
ax1.set_xticks(ticks=list(range(0,110,10)), minor=False)
plt.title('Data Representation: Dog Breed vs. Image Count')
plt.xlabel('Image Count')
plt.ylabel('Dog Breed')
plt.show();
Visualization Reflection:¶
The Alaskan Malamute has the most images at ~96 counted.
The Xoloitzcuintli has fewest images at ~33 counted.
The median Anatolian_shepherd_dog has ~63 images counted.
Data Exloration and Visualization, Image File Size by Breed¶
# Data Representation of Image File Sizes
# Find median file sizes by breed and data set
train_size = np.asarray([os.stat(i).st_size for i in train_files])
valid_size = np.asarray([os.stat(i).st_size for i in valid_files])
test_size = np.asarray([os.stat(i).st_size for i in test_files])
byt_trn = []
byt_val = []
byt_tst = []
for name in nnames:
#train
trn_szs = np.where(train_breeds==name, train_size, 0)
trn_szs_m = np.ma.masked_equal(trn_szs, 0)
byt_trn.append(np.ma.median(trn_szs_m,axis=0))
#validate
val_szs = np.where(valid_breeds==name, valid_size, 0)
val_szs_m = np.ma.masked_equal(val_szs, 0)
byt_val.append(np.ma.median(val_szs_m,axis=0))
#test
tst_szs = np.where(test_breeds==name, test_size, 0)
tst_szs_m = np.ma.masked_equal(tst_szs, 0)
byt_tst.append(np.ma.median(tst_szs_m,axis=0))
byt_sum = [sum(x) for x in zip(byt_trn, byt_val, byt_tst)]
df_bytes = pd.DataFrame.from_dict({'name': nnames,
'train': byt_trn,
'valid': byt_val,
'test': byt_tst,
'total': byt_sum,
})
df_bytes = df_bytes.sort_values(by='total', ascending=False)
df_bytes = df_bytes.drop(['total'], axis=1)
df_bytes = df_bytes.reset_index(drop=True)
df_bytes.head()
# Median Image Size Dog Breed
df_bytes.iloc[66]['name']
# Median Image File Byte Size Statistics
df_bytes.describe()
Image size Reflection:¶
The median of the medians seems similar at ~82 kb.
The smalest image is 20 kb, the largest 325 kb.
The total median bytes varies from 110 kb to 526 kb.
Equalizing data disparity perfectly seems unlikely.
Fundamentally, the data set is imbalanced.
We expect some breeds to test better than others.
fig, ax1 = plt.subplots(1,1, figsize=(7, 12), dpi=120)
df_bytes.plot.barh(x = 'name', stacked=True,
fontsize = 6, ax = ax1,
alpha = 0.5)
plt.title('Median Image File Size by Dog Breed')
plt.ylabel('Dog Breed')
plt.xlabel('Bytes')
plt.show();
Visualization Reflection:¶
The images for the Yorshire Terrier tend to be the largest.
The images for the American Water Spaniel seem to be the smallest.
The French Bulldog sits in the middle of the pack at ~270 kb
sum total for train validate and test median image sizes.
We expect the image preprocessing to even this out to some extent.
The shape of the preprocessing image tensors is (224,224,3).
As a large jpeg image file of many pixels is converted to a tensor,
we expect to lose information in a way that balances the data.
Data Exloration and Visualization, Image File Shape¶
# build a shape dataframe
def findshape(files,breeds,dset):
'''
import image files, breeds, dataset nickname
load image and parse shape into categories
return dataframe
'''
names = []
heights = []
widths = []
aspects = []
dsets = []
for f,b in zip(files, breeds):
try:
fdata = mplim.imread(f)
h,w,d = fdata.shape
names.append(b)
heights.append(h)
widths.append(w)
aspects.append(round(h/w,2))
dsets.append(dset)
except:
print('Error loading: ' + f)
df_shape = pd.DataFrame.from_dict({'name': names,
'height': heights,
'width': widths,
'aspect': aspects,
'dataset': dsets,
})
return df_shape
dfs_train = findshape(train_files, train_breeds,'train')
dfs_valid = findshape(valid_files, valid_breeds,'valid')
dfx_test = findshape(test_files, test_breeds,'test')
df_shape = pd.concat([dfs_train,dfs_valid,dfx_test], ignore_index=True)
df_shape.info()
# Show Statistics on Shape
df_shape.describe()
Image shape Reflection:¶
The average Height to Width Aspect ratio is 1.
That is, the average image is square shape.
The tallest image has a height ~2.5 times its width.
The widest image has a height ~0.3 times its width.
The median aspect ratio is less than one.
This suggests images tend to be more landscape format,
than portrait format.
However, there could be situations of square images nested
inside rectangular format files, where outlier pixels
are not contributing information of interest.
# Plot Aspect Ratio
fig, ax = plt.subplots(1,1, figsize=(8, 5), dpi=120)
bins = np.arange(0.2,2.8, 0.2)
ax = sb.distplot(df_shape.aspect, bins=bins, hist_kws= dict(edgecolor="k"))
ax.set(xlabel='Aspect Ratio', ylabel='Frequency')
ax.set_xticks(bins[:-1]+0.1)
ax.set_title('Image File Shapes: Height to Width Aspect Ratios')
plt.show()
Visualization Reflection:¶
There is a right skew typical when the median or mode is less than the mean.
As the image tensor is square, some pixel data may be cropped or distorted.
A preprocessing step of scaling images on image load relates to this.
Step 1: Detect Humans¶
We use OpenCV's implementation of Haar feature-based cascade classifiers to detect human faces in images. OpenCV provides many pre-trained face detectors, stored as XML files on github. We have downloaded one of these detectors and stored it in the haarcascades directory.
In the next code cell, we demonstrate how to use this detector to find human faces in a sample image.
import cv2
import matplotlib.pyplot as plt
%matplotlib inline
# extract pre-trained face detector
face_cascade = cv2.CascadeClassifier('haarcascades/haarcascade_frontalface_alt.xml')
# load color (BGR) image
img = cv2.imread(human_files[11])
# convert BGR image to grayscale
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# find faces in image
faces = face_cascade.detectMultiScale(gray)
# print number of faces detected in the image
print('Number of faces detected:', len(faces))
# get bounding box for each detected face
for (x,y,w,h) in faces:
# add bounding box to color image
cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)
# convert BGR image to RGB for plotting
cv_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# display the image, along with bounding box
plt.axis('off')
plt.imshow(cv_rgb)
plt.show()
Before using any of the face detectors, it is standard procedure to convert the images to grayscale. The detectMultiScale function executes the classifier stored in face_cascade and takes the grayscale image as a parameter.
In the above code, faces is a numpy array of detected faces, where each row corresponds to a detected face. Each detected face is a 1D array with four entries that specifies the bounding box of the detected face. The first two entries in the array (extracted in the above code as x and y) specify the horizontal and vertical positions of the top left corner of the bounding box. The last two entries in the array (extracted here as w and h) specify the width and height of the box.
Write a Human Face Detector¶
We can use this procedure to write a function that returns True if a human face is detected in an image and False otherwise. This function, aptly named face_detector, takes a string-valued file path to an image as input and appears in the code block below.
# returns "True" if face is detected in image stored at img_path
def face_detector(img_path):
'''
import the image path
convert path to image object
convert image to grayscale
classify the image as containing a face or not
return boolean result of classification
'''
img = cv2.imread(img_path)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
faces = face_cascade.detectMultiScale(gray)
return len(faces) > 0
(IMPLEMENTATION) Assess the Human Face Detector¶
Question 1: Use the code cell below to test the performance of the face_detector function.
- What percentage of the first 100 images in
human_fileshave a detected human face? - What percentage of the first 100 images in
dog_fileshave a detected human face?
Ideally, we would like 100% of human images with a detected face and 0% of dog images with a detected face. You will see that our algorithm falls short of this goal, but still gives acceptable performance. We extract the file paths for the first 100 images from each of the datasets and store them in the numpy arrays human_files_short and dog_files_short.
Answer:
- The percentage of human faces found in the first 100 human pictures is 100.
- The percentage of human faces found in the first 100 dog pictures is 11.
human_files_short = human_files[:100]
dog_files_short = train_files[:100]
# Do NOT modify the code above this line.
## TODO: Test the performance of the face_detector algorithm
## on the images in human_files_short and dog_files_short.
percent_hfhp = np.sum(list(map(face_detector,human_files_short)))
percent_hfdp = np.sum(list(map(face_detector,dog_files_short)))
print('percent human faces found in human pics = ', percent_hfhp )
print('percent human faces found in dog pics = ', percent_hfdp )
Question 2: This algorithmic choice necessitates that we communicate to the user that we accept human images only when they provide a clear view of a face (otherwise, we risk having unneccessarily frustrated users!). In your opinion, is this a reasonable expectation to pose on the user? If not, can you think of a way to detect humans in images that does not necessitate an image with a clearly presented face?
Answer:
This seems a reasonable expectation given that the model is trained consistent to this expectation.
For example, a human wears clothes. It is conceivable that some clothes appear dog like in nature.
However, a human face is used here as a differentiating factor that distinguishes a human presence.
We suggest the face detector from OpenCV as a potential way to detect human images in your algorithm, but you are free to explore other approaches, especially approaches that make use of deep learning :). Please use the code cell below to design and test your own face detection algorithm. If you decide to pursue this optional task, report performance on each of the datasets.
## (Optional) TODO: Report the performance of another
## face detection algorithm on the LFW dataset
def face_detector2(img_path):
'''
input image path
convert path to image object
convert object to grayscale
classify based upon refined hyperparameters
return boolean result as containing face or not
'''
img = cv2.imread(img_path)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
faces = face_cascade.detectMultiScale(
gray,
scaleFactor=1.15,
minNeighbors=5,
minSize=(30, 30),
)
return len(faces) > 0
percent_hfhp2 = np.sum(list(map(face_detector2,human_files_short)))
percent_hfdp2 = np.sum(list(map(face_detector2,dog_files_short)))
print('Upgraded Human Face Detection Attempt')
print('percent human faces found in human pics = ', percent_hfhp2 )
print('percent human faces found in dog pics = ', percent_hfdp2 )
Evolved Detector Reflection:¶
The scale factor hyer parameter was changed from 1.1 to 1.5,
this has the effect of scaling the image more to find features.
This degraded the human face performance but improved
the dog pic performance significantly.
The end algo has a hierarchical sequence of looking for dogs,
then if it doesn't find a dog, to look for a human.
In the case it misses a dog, we don't want it to make
a second mistake of finding a human in a dog picture.
Step 2: Detect Dogs¶
In this section, we use a pre-trained ResNet-50 model to detect dogs in images. Our first line of code downloads the ResNet-50 model, along with weights that have been trained on ImageNet, a very large, very popular dataset used for image classification and other vision tasks. ImageNet contains over 10 million URLs, each linking to an image containing an object from one of 1000 categories. Given an image, this pre-trained ResNet-50 model returns a prediction (derived from the available categories in ImageNet) for the object that is contained in the image.
from keras.applications.resnet50 import ResNet50
# define ResNet50 model
ResNet50_model = ResNet50(weights='imagenet')
Pre-process the Data¶
When using TensorFlow as backend, Keras CNNs require a 4D array (which we'll also refer to as a 4D tensor) as input, with shape
$$ (\text{nb_samples}, \text{rows}, \text{columns}, \text{channels}), $$where nb_samples corresponds to the total number of images (or samples), and rows, columns, and channels correspond to the number of rows, columns, and channels for each image, respectively.
The path_to_tensor function below takes a string-valued file path to a color image as input and returns a 4D tensor suitable for supplying to a Keras CNN. The function first loads the image and resizes it to a square image that is $224 \times 224$ pixels. Next, the image is converted to an array, which is then resized to a 4D tensor. In this case, since we are working with color images, each image has three channels. Likewise, since we are processing a single image (or sample), the returned tensor will always have shape
The paths_to_tensor function takes a numpy array of string-valued image paths as input and returns a 4D tensor with shape
Here, nb_samples is the number of samples, or number of images, in the supplied array of image paths. It is best to think of nb_samples as the number of 3D tensors (where each 3D tensor corresponds to a different image) in your dataset!
from keras.preprocessing import image
from tqdm import tqdm
def path_to_tensor(img_path):
'''
input image path
load image
render into compatible tensor
return tensor
'''
# loads RGB image as PIL.Image.Image type
img = image.load_img(img_path, target_size=(224, 224),
interpolation = "bicubic")
# convert PIL.Image.Image type to 3D tensor with shape (224, 224, 3)
x = image.img_to_array(img)
# convert 3D tensor to 4D tensor with shape (1, 224, 224, 3) and return 4D tensor
return np.expand_dims(x, axis=0)
def paths_to_tensor(img_paths):
'''
input list of image paths
render paths to tensors
return vertical stack of tensors
'''
list_of_tensors = [path_to_tensor(img_path) for img_path in tqdm(img_paths)]
return np.vstack(list_of_tensors)
Preprocessing Reflection¶
We saw earlier most images are not square.
The default image load scaling format is nearest.
Potentially, improvement in retaining image fidelity
may be attained in using bilinear or bicubic scaling
within the image load preprocessing step. References:
https://keras.io/api/preprocessing/image/
https://en.wikipedia.org/wiki/Comparison_gallery_of_image_scaling_algorithms
Making Predictions with ResNet-50¶
Getting the 4D tensor ready for ResNet-50, and for any other pre-trained model in Keras, requires some additional processing. First, the RGB image is converted to BGR by reordering the channels. All pre-trained models have the additional normalization step that the mean pixel (expressed in RGB as $[103.939, 116.779, 123.68]$ and calculated from all pixels in all images in ImageNet) must be subtracted from every pixel in each image. This is implemented in the imported function preprocess_input. If you're curious, you can check the code for preprocess_input here.
Now that we have a way to format our image for supplying to ResNet-50, we are now ready to use the model to extract the predictions. This is accomplished with the predict method, which returns an array whose $i$-th entry is the model's predicted probability that the image belongs to the $i$-th ImageNet category. This is implemented in the ResNet50_predict_labels function below.
By taking the argmax of the predicted probability vector, we obtain an integer corresponding to the model's predicted object class, which we can identify with an object category through the use of this dictionary.
from keras.applications.resnet50 import preprocess_input, decode_predictions
def ResNet50_predict_labels(img_path):
'''
input image path
convert path to tensor
preprocess the tensor
make predictions
return most probable prediction
'''
# returns prediction vector for image located at img_path
img = preprocess_input(path_to_tensor(img_path))
return np.argmax(ResNet50_model.predict(img))
Write a Dog Detector¶
While looking at the dictionary, you will notice that the categories corresponding to dogs appear in an uninterrupted sequence and correspond to dictionary keys 151-268, inclusive, to include all categories from 'Chihuahua' to 'Mexican hairless'. Thus, in order to check to see if an image is predicted to contain a dog by the pre-trained ResNet-50 model, we need only check if the ResNet50_predict_labels function above returns a value between 151 and 268 (inclusive).
We use these ideas to complete the dog_detector function below, which returns True if a dog is detected in an image (and False if not).
### returns "True" if a dog is detected in the image stored at img_path
def dog_detector(img_path):
'''
input image path
retrieve most probable prediction
classify prediction as likely a dog
return boolean result
'''
prediction = ResNet50_predict_labels(img_path)
return ((prediction <= 268) & (prediction >= 151))
(IMPLEMENTATION) Assess the Dog Detector¶
Question 3: Use the code cell below to test the performance of your dog_detector function.
- What percentage of the images in
human_files_shorthave a detected dog? - What percentage of the images in
dog_files_shorthave a detected dog?
Answer:
- The percentage of dogs found in the first 100 human pictures is 0.
- The percentage of dogs found in the first 100 dog pictures is 100.
### TODO: Test the performance of the dog_detector function
### on the images in human_files_short and dog_files_short.
percent_dfhp = np.sum(list(map(dog_detector, human_files_short)))
percent_dfdp = np.sum(list(map(dog_detector, dog_files_short)))
print('percent dog faces found in human pics = ', percent_dfhp)
print('percent dog faces found in dog pics = ', percent_dfdp)
Step 3: Create a CNN to Classify Dog Breeds (from Scratch)¶
Now that we have functions for detecting humans and dogs in images, we need a way to predict breed from images. In this step, you will create a CNN that classifies dog breeds. You must create your CNN from scratch (so, you can't use transfer learning yet!), and you must attain a test accuracy of at least 1%. In Step 5 of this notebook, you will have the opportunity to use transfer learning to create a CNN that attains greatly improved accuracy.
Be careful with adding too many trainable layers! More parameters means longer training, which means you are more likely to need a GPU to accelerate the training process. Thankfully, Keras provides a handy estimate of the time that each epoch is likely to take; you can extrapolate this estimate to figure out how long it will take for your algorithm to train.
We mention that the task of assigning breed to dogs from images is considered exceptionally challenging. To see why, consider that even a human would have great difficulty in distinguishing between a Brittany and a Welsh Springer Spaniel.
| Brittany | Welsh Springer Spaniel |
|---|---|
 |
 |
It is not difficult to find other dog breed pairs with minimal inter-class variation (for instance, Curly-Coated Retrievers and American Water Spaniels).
| Curly-Coated Retriever | American Water Spaniel |
|---|---|
 |
 |
Likewise, recall that labradors come in yellow, chocolate, and black. Your vision-based algorithm will have to conquer this high intra-class variation to determine how to classify all of these different shades as the same breed.
| Yellow Labrador | Chocolate Labrador | Black Labrador |
|---|---|---|
 |
 |
 |
We also mention that random chance presents an exceptionally low bar: setting aside the fact that the classes are slightly imabalanced, a random guess will provide a correct answer roughly 1 in 133 times, which corresponds to an accuracy of less than 1%.
Remember that the practice is far ahead of the theory in deep learning. Experiment with many different architectures, and trust your intuition. And, of course, have fun!
Pre-process the Data¶
We rescale the images by dividing every pixel in every image by 255.
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
# pre-process the data for Keras
train_tensors = paths_to_tensor(train_files).astype('float32')/255
valid_tensors = paths_to_tensor(valid_files).astype('float32')/255
test_tensors = paths_to_tensor(test_files).astype('float32')/255
(IMPLEMENTATION) Model Architecture¶
Create a CNN to classify dog breed. At the end of your code cell block, summarize the layers of your model by executing the line:
model.summary()
We have imported some Python modules to get you started, but feel free to import as many modules as you need. If you end up getting stuck, here's a hint that specifies a model that trains relatively fast on CPU and attains >1% test accuracy in 5 epochs:
Question 4: Outline the steps you took to get to your final CNN architecture and your reasoning at each step. If you chose to use the hinted architecture above, describe why you think that CNN architecture should work well for the image classification task.
Answer:
Network Outline:
- Squeeze out spatial information: six layers consist of a 2D convolutional layer followed by a max pooling layer. This >provides spatial invariance of dog breed features in identifying the dog breed, meaning the key features can move around >within the picture frame without the dog breed features escaping the classifier.
- Flatten 2D inputs to 1D output: The 2D output of the last pooling layer is flattened into 1D in the 7th layer, as a bridge >toward a 1D output format.
- Neurally link the output: Two Dense layers separated by a Dropout layer are intended to map the convolution layer output >into the classification categories. The Dense layers left to their own device are apt to overtrain the classifier and the >dropout layers act as counter measures against an over training effect, where the network might memorize the train data in a >way it cannot validate well on the test data.
- Provide Output Probabilities: The last dense layer contains one entry per target class with a Softmax activation function >that provides output probabilities for the target classes.
These convolution layers with 3x3 masks in strides of 2 followed by
a pooling layer, were found to be useful in the Resnet architecture for
a similar image classification task.
This is why I think this should work well, however the number of layers
is not sufficient to perform as well as Resnet.
from keras.layers import Conv2D, MaxPooling2D, GlobalAveragePooling2D
from keras.layers import Dropout, Flatten, Dense
from keras.models import Sequential
model = Sequential()
### TODO: Define your architecture.
model.add(Conv2D(filters=32, kernel_size=3, strides=2,
padding='same', activation='relu',input_shape=(224, 224, 3)))
model.add(MaxPooling2D(pool_size=2))
model.add(Conv2D(filters=64, kernel_size=3, strides=2,
padding='same', activation='relu'))
model.add(MaxPooling2D(pool_size=2))
model.add(Conv2D(filters=128, kernel_size=3, strides=2,
padding='same', activation='relu'))
model.add(MaxPooling2D(pool_size=2))
model.add(Flatten())
model.add(Dropout(0.2))
model.add(Dense(256, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(133, activation='softmax'))
model.summary()
Compile the Model¶
model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])
(IMPLEMENTATION) Train the Model¶
Train your model in the code cell below. Use model checkpointing to save the model that attains the best validation loss.
You are welcome to augment the training data, but this is not a requirement.
from keras.callbacks import ModelCheckpoint
### TODO: specify the number of epochs that you would like to use to train the model.
epochs = 15
### Do NOT modify the code below this line.
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.from_scratch.hdf5',
verbose=1, save_best_only=True)
model.fit(train_tensors, train_targets,
validation_data=(valid_tensors, valid_targets),
epochs=epochs, batch_size=20, callbacks=[checkpointer], verbose=1)
Load the Model with the Best Validation Loss¶
model.load_weights('saved_models/weights.best.from_scratch.hdf5')
Test the Model¶
Try out your model on the test dataset of dog images. Ensure that your test accuracy is greater than 1%.
# get index of predicted dog breed for each image in test set
dog_breed_predictions = [np.argmax(model.predict(np.expand_dims(tensor, axis=0))) for tensor in test_tensors]
# report test accuracy
test_accuracy = 100*np.sum(np.array(dog_breed_predictions)==np.argmax(test_targets, axis=1))/len(dog_breed_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
# Print classification report on test data
print(classification_report(np.argmax(test_targets, axis=1),
dog_breed_predictions,
target_names=nnames))
Simple Network Result Evaluation:¶
The summary yields a warning of many classes without predictions.
The report then sets Precision and F1 to zero, but it may be misleading.
Given more opportunity, predicted samples may occur lifting these positive.
Regardless, the accuracy found of 12% indicates room for improvement.
The Alaskan Malamute and Basset Hound revealed high image count representation.
These two breeds seem to perform better than the other extreme.
The Xoloitzcuintli and Plott have low image counts and zero scores here.
bottleneck_features = np.load('bottleneck_features/DogVGG16Data.npz')
train_VGG16 = bottleneck_features['train']
valid_VGG16 = bottleneck_features['valid']
test_VGG16 = bottleneck_features['test']
Model Architecture¶
The model uses the the pre-trained VGG-16 model as a fixed feature extractor, where the last convolutional output of VGG-16 is fed as input to our model. We only add a global average pooling layer and a fully connected layer, where the latter contains one node for each dog category and is equipped with a softmax.
VGG16_model = Sequential()
VGG16_model.add(GlobalAveragePooling2D(input_shape=train_VGG16.shape[1:]))
VGG16_model.add(Dense(133, activation='softmax'))
VGG16_model.summary()
Compile the Model¶
VGG16_model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
Train the Model¶
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.VGG16.hdf5',
verbose=1, save_best_only=True)
VGG16_model.fit(train_VGG16, train_targets,
validation_data=(valid_VGG16, valid_targets),
epochs=90, batch_size=30, callbacks=[checkpointer], verbose=1)
Load the Model with the Best Validation Loss¶
VGG16_model.load_weights('saved_models/weights.best.VGG16.hdf5')
Test the Model¶
Now, we can use the CNN to test how well it identifies breed within our test dataset of dog images. We print the test accuracy below.
# get index of predicted dog breed for each image in test set
VGG16_predictions = [np.argmax(VGG16_model.predict(np.expand_dims(feature, axis=0))) for feature in test_VGG16]
# report test accuracy
test_accuracy = 100*np.sum(np.array(VGG16_predictions)==np.argmax(test_targets, axis=1))/len(VGG16_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
# Print classification report on test data
print(classification_report(np.argmax(test_targets, axis=1),
VGG16_predictions,
target_names=nnames))
VGG16 Network Result Evaluation:¶
The summary yields a warning again, of many classes without predictions.
The report then sets Precision and F1 to zero, but it may be misleading.
In this case the accuracy improved up to 51 percent.
The Alaskan Malamute and Basset Hound revealed high image count representation.
These two breeds however got some big fat zeros here.
By contrast, the Xoloitzcuintli and Plott have low image counts and higher scores.
Possibly the VGG16 pre-train agrees with the bicubic imported image tensors.
Predict Dog Breed with the Model¶
from extract_bottleneck_features import *
def VGG16_predict_breed(img_path):
'''
input image path
convert image to tensor
retrieve output of pre-trained network, aka bottleneck
make prediction from full connected network
return prediction of dog breed
'''
# extract bottleneck features
bottleneck_feature = extract_VGG16(path_to_tensor(img_path))
# obtain predicted vector
predicted_vector = VGG16_model.predict(bottleneck_feature)
# return dog breed that is predicted by the model
return dog_names[np.argmax(predicted_vector)]
Step 5: Create a CNN to Classify Dog Breeds (using Transfer Learning)¶
You will now use transfer learning to create a CNN that can identify dog breed from images. Your CNN must attain at least 60% accuracy on the test set.
In Step 4, we used transfer learning to create a CNN using VGG-16 bottleneck features. In this section, you must use the bottleneck features from a different pre-trained model. To make things easier for you, we have pre-computed the features for all of the networks that are currently available in Keras:
- VGG-19 bottleneck features
- ResNet-50 bottleneck features
- Inception bottleneck features
- Xception bottleneck features
The files are encoded as such:
Dog{network}Data.npz
where {network}, in the above filename, can be one of VGG19, Resnet50, InceptionV3, or Xception. Pick one of the above architectures, download the corresponding bottleneck features, and store the downloaded file in the bottleneck_features/ folder in the repository.
(IMPLEMENTATION) Obtain Bottleneck Features¶
In the code block below, extract the bottleneck features corresponding to the train, test, and validation sets by running the following:
bottleneck_features = np.load('bottleneck_features/Dog{network}Data.npz')
train_{network} = bottleneck_features['train']
valid_{network} = bottleneck_features['valid']
test_{network} = bottleneck_features['test']### TODO: Obtain bottleneck features from another pre-trained CNN.
bottleneck_features = np.load('bottleneck_features/DogResnet50Data.npz')
train_RS50 = bottleneck_features['train']
valid_RS50 = bottleneck_features['valid']
test_RS50 = bottleneck_features['test']
(IMPLEMENTATION) Model Architecture¶
Create a CNN to classify dog breed. At the end of your code cell block, summarize the layers of your model by executing the line:
<your model's name>.summary()
Question 5: Outline the steps you took to get to your final CNN architecture and your reasoning at each step. Describe why you think the architecture is suitable for the current problem.
Answer:
Transfer learning adopts a pre-trained neural network and adapts it to a specific data set.
- Interface to the pre-trained network. A fully connected layer takes the existing
outputs and bridges them over to a fully connected dense layer. Weights assigned
to this interface are trained to complement optimal bridging. Here the number
of nodes matches the number of nodes on the pre-trained network.Convey an output correponding to target classes. A fully connected dense layer
provides the probability outputs corresponding to the target classes.
Here, the number of nodes matches the number of output classes of interest.This seems an example of interfacing a relatively small data set of similar images
that complement fine tuning. That is, the images are not so dissimilar than those
that Resnet50 was originally trained upon, and the data set is small by comparison.Resnet50 was trained upon Image Net 2012 classification dataset
that contains 1000 classes, trained on 1.28M images, validated on 50K images,
and tested on 100K images. Approximately 120 classes of Resnet were dog breeds.By comarison, we have ~7K train images, and ~800 each validation and test images,
and 133 classes of dog breeds which is relatively small to Resnet's origin data.In the case of a small set of similar data, the network is already trained
in a way that is relevant and complements the data. It simply needs an output
interface tailored to the special cases that the data is being used for.
As such, simply the end of the pre-trained network is trained using these
two fully connected layers.ResNet Paper Reference: https://arxiv.org/abs/1512.03385
### TODO: Define your architecture.
RS50_model = Sequential()
RS50_model.add(GlobalAveragePooling2D(input_shape=train_RS50.shape[1:]))
RS50_model.add(Dense(133, activation='softmax'))
RS50_model.summary()
(IMPLEMENTATION) Compile the Model¶
### TODO: Compile the model.
RS50_model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
(IMPLEMENTATION) Train the Model¶
Train your model in the code cell below. Use model checkpointing to save the model that attains the best validation loss.
You are welcome to augment the training data, but this is not a requirement.
### TODO: Train the model.
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.Resnet50.hdf5',
verbose=1,save_best_only=True)
RS50_model.fit(train_RS50, train_targets,
validation_data=(valid_RS50, valid_targets),
epochs=20, batch_size=30, callbacks=[checkpointer], verbose=1)
(IMPLEMENTATION) Load the Model with the Best Validation Loss¶
### TODO: Load the model weights with the best validation loss.
RS50_model.load_weights('saved_models/weights.best.Resnet50.hdf5')
(IMPLEMENTATION) Test the Model¶
Try out your model on the test dataset of dog images. Ensure that your test accuracy is greater than 60%.
### TODO: Calculate classification accuracy on the test dataset.
RS50_predictions = [np.argmax(RS50_model.predict(np.expand_dims(feature, axis=0))) for feature in test_RS50]
# report test accuracy
test_accuracy = 100*np.sum(np.array(RS50_predictions)==np.argmax(test_targets, axis=1))/len(RS50_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
# Print classification report on test data
print(classification_report(np.argmax(test_targets, axis=1),
RS50_predictions,
target_names=nnames))
RS50 Network Result Evaluation:¶
The summary yields no warnings, which implies predictions in all classes.
In this case the accuracy improved up to 82% on the test cases.
This agrees well with the validation set suggesting it isn't overtrained.
The avg precision is 85% indicates it is more relevant than irrelevant,
in other words, there are 15% false positives.
The avg recall is 82% indicates the percentage of relevant caught is high,
in other words, there are 18% false negatives.
Returning to accuracy, slightly better than one out of five times,
it slips up on identifying the correct breed of dog depicted within the image.
This may be partially related to inconsistencies in photography and data handling.
Justification Reflection¶
A sequence of models was built of incrementally improving quality.
The simple model resulted in the worst performance, 11% accurate.
The Pre-Trained VGG16 tailored to these dogs improved, 51% accurate.
The Pre-Trained RS50, which contained dog breed classes to begin with,
was tailored to these dogs and resulted in best performance,
82% accurate.Unseen to the reader may be the essense of training time.
That is, the RS50 network took immense effort to build.
However, to tailor it to this aim entailed an effort that seems
comparable to the simple network, however perfoms 8 times better.
For this reason, it seems that transfer learning approaches offer
fast tracking advantage in arriving at classification solutions.
(IMPLEMENTATION) Predict Dog Breed with the Model¶
Write a function that takes an image path as input and returns the dog breed (Affenpinscher, Afghan_hound, etc) that is predicted by your model.
Similar to the analogous function in Step 5, your function should have three steps:
- Extract the bottleneck features corresponding to the chosen CNN model.
- Supply the bottleneck features as input to the model to return the predicted vector. Note that the argmax of this prediction vector gives the index of the predicted dog breed.
- Use the
dog_namesarray defined in Step 0 of this notebook to return the corresponding breed.
The functions to extract the bottleneck features can be found in extract_bottleneck_features.py, and they have been imported in an earlier code cell. To obtain the bottleneck features corresponding to your chosen CNN architecture, you need to use the function
extract_{network}
where {network}, in the above filename, should be one of VGG19, Resnet50, InceptionV3, or Xception.
### TODO: Write a function that takes a path to an image as input
### and returns the dog breed that is predicted by the model.
from extract_bottleneck_features import *
try:
# extract bottleneck features
bottleneck_feature = extract_Resnet50()
# obtain predicted vector
predicted_vector = RS50_model.predict(bottleneck_feature)
except:
pass
def RS50_predict_breed(img_path):
'''
input image path
convert image to tensor
retrieve output of pre-trained network, aka bottleneck
make prediction from full connected network
return prediction of dog breed
'''
# extract bottleneck features
bottleneck_feature = extract_Resnet50(path_to_tensor(img_path))
# obtain predicted vector
predicted_vector = RS50_model.predict(bottleneck_feature)
# return dog breed that is predicted by the model
return dog_names[np.argmax(predicted_vector)]
Step 6: Write your Algorithm¶
Write an algorithm that accepts a file path to an image and first determines whether the image contains a human, dog, or neither. Then,
- if a dog is detected in the image, return the predicted breed.
- if a human is detected in the image, return the resembling dog breed.
- if neither is detected in the image, provide output that indicates an error.
You are welcome to write your own functions for detecting humans and dogs in images, but feel free to use the face_detector and dog_detector functions developed above. You are required to use your CNN from Step 5 to predict dog breed.
A sample image and output for our algorithm is provided below, but feel free to design your own user experience!
This photo looks like an Afghan Hound.
(IMPLEMENTATION) Write your Algorithm¶
def pic_display(path):
'''
load image path
convert to RGB
display in RGB
'''
img = cv2.imread(path)
cv_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
plt.imshow(cv_rgb)
plt.axis('off')
plt.show()
def picture_sort(path):
'''
import image path
display image
sort image by dog or human
if dog, figure out the breed
print sort decision
'''
# display
pic_display(path)
# sort class and reveal
dog = dog_detector(path)
human = face_detector2(path)
if dog:
print('A dog detected!')
prob_breed = RS50_predict_breed(path)
per_loc = prob_breed.find('.')+1
print('Possible breed: '+ prob_breed[per_loc:])
elif human:
print('A human face detected!')
prob_breed = RS50_predict_breed(path)
per_loc = prob_breed.find('.')+1
print('May share resemblance with: '+ prob_breed[per_loc:])
else:
print('Neither a dog nor a human face was detected.')
print('----------------------------------------------')
print()
Step 7: Test Your Algorithm¶
In this section, you will take your new algorithm for a spin! What kind of dog does the algorithm think that you look like? If you have a dog, does it predict your dog's breed accurately? If you have a cat, does it mistakenly think that your cat is a dog?
(IMPLEMENTATION) Test Your Algorithm on Sample Images!¶
Test your algorithm at least six images on your computer. Feel free to use any images you like. Use at least two human and two dog images.
Question 6: Is the output better than you expected :) ? Or worse :( ? Provide at least three possible points of improvement for your algorithm.
Answer:
- The output is better than expected.
Only one error was observed out of eleven attempts.
The error occurred in what seems like a picture of a dog picture.
This was recognized as neither a dog nor a human.
In taking pictures of pictures, the picture fidelity can degrade.
Potential errors made in dog breed classification seem questionable.
There's many cross breeds and DNA analysis is needed for accuracy.- Potential improvement in the algorithm include:
A. Detection of multiple dogs or humans or mix in pictures.
B. Expansion of classes to include other animal types, i.e. fox, wolf.
C. Picture fidelity screening to prevent errors.
## TODO: Execute your algorithm from Step 6 on
## at least 6 images on your computer.
## Feel free to use as many code cells as needed.
import os
import glob
import cv2
cwd = os.getcwd()
img_folder = cwd + "/images/*"
pics = glob.glob(img_folder)
for path in pics:
picture_sort(path)
Conclusion¶
In this project we tried to solve an image classification problem.
Given an image, if it contains a dog then answer what breed it might be.
Otherwise, if it contains a face then answer what dog breed shares resemblance.
To tackle this problem, we pursued a supervised learing approach.
To do that, we jumped off the shoulders of giants: beloved data curators at Udacity.
The dataset utilized contained over 8000 dog images and 13000 human images.
In taking a close look at dog data, we discovered the image count per breed
varies by a factor of three, the file size varies by a factor of 5, and the
aspect ratio of the pictures varies from wide landscapes to tall portraits.
Using a divide and conquer approach, first a human face detector was built.
This function input an image and returned a true false reply if a face was found.
At the heart of the human face detector works the OpenCV face cascade module.
Furthermore, we built a dog detector using ResNet-50 with imagenet weights.
Both the dog detector and face detector were found to be highly accurate, ~100%.
From there, focus was applied to three classifiers for distinguishing 133 breeds.
A simple Convolutional Neural Network, a Transfer learned VGG16, and a transfer
learned ResNet50. These resulted in accuracies of 11%, 54%, and 81% repectively.
Finally the face and dog detector were combined with the breed classifier
in a way that addressed the crux of the problem, given an image - provide answers.The successful results revealed by the output demonstrates a classification
approach that likely has advantage over others for compatible classification schemes.
Here a transfer learning approach attained a good result in a relatively short time.
I found it interesting that GPU training appears to progress faster than CPU alone.
The indication of progress seems to affirm problems are solved sooner than later.
I also found it interesting that this is one of the most popular projects.
I attribute much of the success and popularity to the dataset, thanks for that.
There seem many blogs featuring this topic, as if man's best friend came to the rescue.One improvement idea for this project is based upon the idea garbage in causes
garbage out. That is, the images are screened for some degrees of quality that
are not meant to trick the network. Making a network that classifies well using
clean data is a step toward making a network that works with dirty data.
Potentially an image fidelity classifier is put on the front end to weed out
images that recklessly degrade performance. Another improvement idea relates
to standardizing image input formats, where variation in image sizes are regularized
before undergoing similar scaling routines into tensors. In this case the image
shapes were found widely varying and the scaling to tensor somewhat mysterious,
where imposing data regularization in the interest of reducing variance may improve
classification noise levels. Lastly, full stack implementation of this onto a
mobile phone app, seems beyond the grasp of this notebook - but what if it wasn't?
It would improve the project to deploy on a mobile phone in a way that puts the
solution of this problem into the hands of dog lovers the world over.
Additonal inference routines¶
import pickle
# pickle dog names
pickle.dump(dog_names, open("dognames.pkl","wb"))
### TODO: Define your architecture.
RS50_infer = Sequential()
RS50_infer.add(GlobalAveragePooling2D(input_shape=(1,1,2048)))
RS50_infer.add(Dense(133, activation='softmax'))
RS50_infer.load_weights('saved_models/weights.best.Resnet50.hdf5')
def RS50_infer_predict(img_path):
'''
input image path
convert image to tensor
retrieve output of pre-trained network, aka bottleneck
make prediction from full connected network
return prediction of dog breed
'''
# extract bottleneck features
bottleneck_feature = extract_Resnet50(path_to_tensor(img_path))
# obtain predicted vector
predicted_vector = RS50_infer.predict(bottleneck_feature)
# return dog breed that is predicted by the model
return dog_names[np.argmax(predicted_vector)]
def infer_sort(path):
'''
import image path
display image
sort image by dog or human
if dog, figure out the breed
print sort decision
'''
# display
pic_display(path)
# sort class and reveal
dog = dog_detector(path)
human = face_detector2(path)
if dog:
print('A dog detected!')
prob_breed = RS50_infer_predict(path)
per_loc = prob_breed.find('.')+1
print('Possible breed: '+ prob_breed[per_loc:])
elif human:
print('A human face detected!')
prob_breed = RS50_infer_predict(path)
per_loc = prob_breed.find('.')+1
print('May share resemblance with: '+ prob_breed[per_loc:])
else:
print('Error: Neither dog nor human face detected.')
print('----------------------------------------------')
print()
# images uploaded from phone
cwd = os.getcwd()
img_folder = cwd + "/phone/*"
pics = glob.glob(img_folder)
for path in pics:
infer_sort(path)
